Article · Buying Decisions

Vendor FDE vs independent FDE: how to decide who you work with

Arkatai 8 min

A forward-deployed engineer from a model maker and an independent FDE can, on paper, do the same work: sit inside your operation, connect the system and take it to production. The title is identical; the mandate is not. The vendor’s FDE has a goal of deploying the stack their company sells. The independent one optimises your operation and preserves the ability to swap the model. The difference is not one of good faith but of incentives and architecture, and you pay it or bank it on the day you want to change model provider.

If you are not yet clear on what this role actually does, read what a forward-deployed engineer is first. Here I take the role as given and focus on a single question: who they work for.

The same title, two mandates

A model maker who puts engineers in the field does so, among other reasons, to get its technology into production and keep it there. That is legitimate and often useful: nobody knows a model’s capabilities and limits better than whoever built it. But that engineer’s success includes adoption of their ecosystem. When a design decision comes up with two equally valid paths, the one that runs through their own primitives gets a nudge that comes not from the operation but from the mandate.

An independent FDE does not sell a model. Their success is measured by your process working at the agreed quality, cost and risk, and by your still being able to change model tomorrow. They treat models as replaceable components and choose them per task on the basis of evaluations. That is the position I hold at Arkatai, and the reason I stay agnostic about the model provider.

The difference, dimension by dimension

These are the five dimensions where the mandate changes the result, even though the title is the same.

DimensionVendor FDEIndependent FDE
MandateGet the vendor’s stack into productionMake the operation work at agreed levels
What it deploysIts model and primitives as the baseThe platform; the model is interchangeable
What it optimisesAdoption and fit into its ecosystemQuality, cost, latency and process risk
Model choiceIts own model by defaultPer task, based on evaluations
When the model changesHigh friction: logic lives in its primitivesRe-evaluate, not rebuild the process

The table does not say one path is right and the other wrong. It says you should read the mandate before signing, because it determines what stays inside your operation when the engineer leaves.

Why it is an architecture problem, not bad faith

The point boards most often misread is simple. You do not need to assume bad intent from the vendor for the risk to be real. The risk is architectural. If, during the deployment, your rules, controls, traces and integrations end up encoded inside one lab’s specific primitives, changing model stops being a contained technical decision and becomes a rewrite of the process. Nobody deceived you. The system was simply built tied to a provider, and now the cost of leaving is high.

The architecture that protects your freedom does the opposite: it keeps the process, its rules, its controls and its trace outside the model, so the model is a connected component and not the foundation. With that separation, changing provider means re-evaluating quality, cost and latency, but not redoing the operation. It is the same logic of treating models as a utility that I explain in service as software: you buy a result, not a commitment to a specific lab.

Your rulesprocess policyYour controlsapproval and escalationYour traceevery decision loggedIntegrationsERP · CRM · emailOrchestrationsteps and retriesModelutility, by APIswappable
The architecture that protects your freedom: your rules, controls, trace and integrations live outside the model, which stays a swappable layer; changing provider is a re-evaluation, not rebuilding the operation.

What happens the day the model changes

Models change on their own. New versions ship, prices drop, a better one appears for your specific task, or your data policy forces you to move part of the work to another provider. That is the day the difference between the two mandates shows.

With an operation built on the vendor’s primitives, migration is a project: you have to find where the logic that assumed that model lived and rebuild it. With an operation where the model is separated from the process, migration is a re-evaluation: you run your test cases again with the candidate model, compare quality and cost, and decide. The ability to make that comparison per task is exactly what I cover in choosing the LLM for each task. Without it, the model choice is effectively made by whoever deployed your system.

Who answers once the engineer leaves

The mandate question has a second half that almost never gets signed, and it is whose work it is once the engineer moves to another project. An FDE is not a rented programmer, so their departure should not leave a hole. It leaves one when the only thing written down was their judgement.

With a vendor FDE, the party you talk to the day the process fails in production is still that vendor’s support, and what that support covers is their model and its availability, not the outcome of your operation. It is an awkward distinction in a board meeting, because the engineer who sat with you for six months was not the one signing the service level. With a managed operation, the party that answers is the one running the process, and it answers for quality, cost, time and exceptions, which is the split I describe in AI managed services. Who pays once the failure has already happened is a different question and I handle it separately in who pays when an AI agent makes a mistake.

Whose answer it is when it fails →Seconded engineeryou answer for the outcomeVendor FDEanswers for its modelManaged operationanswers for the outcome
Three arrangements with the same engineer in your office and three different answers to “who signs that the process still hits the agreed level” once that engineer is gone.

What makes the engineer’s departure a non-event is that their work is written down and belongs to you. The operating architecture specific to your process, meaning the rules, the exceptions, the controls, the data contracts, the integration specifications and the evaluation criteria, has to be documented in your name and handed over up to date, whoever supplied the engineer. Not the platform, not the platform’s code, but that package yes. Without it you can neither hire another implementer nor sustain a decision in front of an auditor, and the knowledge left with a person who is no longer there.

The question I close this part with in a meeting is always the same. If this engineer moves to another project tomorrow, who signs that the process still hits the agreed level, and with what document do they sign it. If the answer is a name and not a document, there is no operation yet, there is dependency.

When each one fits

The vendor FDE can be the right choice if your company has already decided to marry an ecosystem and does not expect to leave it: in that case, having the person who knows the model best speeds up the deployment and reduces surprises. What you should not do is make that decision without seeing it: choosing the vendor’s FDE is choosing their model, and it is worth measuring the cost of changing later in advance.

The independent FDE fits when you want to keep the model substitutable, when the process is specific enough that technical neutrality matters, or when you do not want the AI provider choice fixed by whoever installed the system. This question is part of a wider decision about who to work with, which I develop in in-house team, consultancy or forward-deployed engineers, and of the underlying decision to buy, build or contract, in buy vs build enterprise AI. The concrete criteria for comparing offers are in how to choose an AI agent provider, and the full map of the terrain, in AI agents for business.

Frequently Asked Questions

What is the difference between a vendor FDE and an independent one?

The mandate. The vendor FDE aims to take into production the model and primitives their company sells, and their success includes adoption of that ecosystem. The independent FDE optimises your operation and treats the model as a replaceable component chosen per task on the basis of evaluations.

Are you implying vendor FDEs act in bad faith?

No. The issue is incentives and architecture, not good faith. A good vendor FDE can do excellent work. What changes is where their mandate pushes and what exit cost your system carries if the logic ends up tied to one specific provider.

Why does being able to change model matter?

Because models change in price, quality and availability constantly, and your data policy may force you to move work. If the process is separated from the model, changing is a re-evaluation. If it is tied to a lab’s primitives, changing means rebuilding the operation.

What happens when the forward-deployed engineer leaves?

It depends on what they wrote down and whose name it is in. If the rules, exceptions, controls, data contracts and evaluation criteria of your process are documented and yours, their departure is a handover. If the only thing there was their judgement in their head, it is a loss of operating capability, and it shows at the first model change or the first odd case.

Who answers for the outcome if the model vendor supplied the FDE?

You do, in practice. The vendor’s support answers for its model and its availability, not for your process hitting the agreed level, and the engineer who sat in your office was not signing that level. If you want somebody to answer for the outcome of the process, that gets contracted as an operation and written into service levels, not inferred from having a good engineer nearby.

When would I choose the vendor FDE?

When the company has already decided to commit to a specific ecosystem and does not expect to leave it. In that case, whoever knows the model best speeds up the deployment. The decision should be made with eyes open: choosing that FDE is, in practice, choosing their model.